src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp

/*
 * Copyright (c) 2022 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include "ClTemplateElementwiseBinary.h"

#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h"
#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.h"

#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "src/core/helpers/WindowHelpers.h"

#include "support/StringSupport.h"

namespace arm_compute
{
namespace experimental
{
namespace dynamic_fusion
{
constexpr unsigned int vector_size_byte_opencl = 16;

ClTemplateElementwiseBinary::ClTemplateElementwiseBinary(ComponentId                      id,
                                                         const ArgumentPack<ITensorInfo> &tensors,
                                                         const Attributes                &attributes)
    : IGpuTemplateComponentWriter{ id, tensors },
      _lhs{},
      _rhs{},
      _dst{},
      _attributes{ attributes }
{
    _lhs = this->tensors().get_const_tensor(TensorType::ACL_SRC_0);
    _rhs = this->tensors().get_const_tensor(TensorType::ACL_SRC_1);
    _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0);
    ARM_COMPUTE_ERROR_ON_NULLPTR(_lhs, _rhs, _dst);
}

std::string ClTemplateElementwiseBinary::get_name() const
{
    return "elementwise_binary";
}

std::string ClTemplateElementwiseBinary::get_component_code(const ComponentGroup &comp_group) const
{
    ARM_COMPUTE_UNUSED(comp_group);
    std::string code;
    const bool  is_broadcast = _lhs->tensor_shape() != _rhs->tensor_shape();
    const bool  is_root      = (comp_group.get_root_component()->id() == this->id());

    if(is_root)
    {
        code =
R"_(
    //------------------ START KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
)_"
    // IN_0(LHS)            {{lhs}}
    // IN_1(RHS)            {{rhs}}
    // OUT(dst, accum)      {{dst}}
    // dst = lhs + rhs (mix-precision, broadcast, boundary aware)
R"_(
    TILE({{DATA_TYPE}}, M0, N0, {{dst}});
    TILE(uint, M0, 1, g_dst_indirect_y);
    {
        TILE({{DATA_TYPE}}, M0, N0, lhs_tile);
        TILE({{DATA_TYPE}}, M0, N0, rhs_tile);
)_"
        // Assuming un-collapsed window
R"_(
        {{lhs}}_offset_first_element_in_bytes += g_ind_2 * {{lhs}}_stride_z;
        {{rhs}}_offset_first_element_in_bytes += g_ind_2 * {{rhs}}_stride_z;

        T_LOAD({{DATA_TYPE}}, M0, N0, BUFFER, {{lhs}}, g_ind_0, g_ind_1, 1, {{lhs}}_stride_y, lhs_tile);
        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{rhs}}, {{rhs_start_ind_0}}, {{rhs_start_ind_1}}, 1, {{rhs}}_stride_y, rhs_tile);
)_";
        if(is_broadcast)
        {
            code +=
R"_(
        T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, lhs_tile, rhs_tile, {{dst}});
)_";
        }
        else
        {
            code +=
R"_(
        T_ELTWISE_{{ELTWISE_OP}}({{DATA_TYPE}}, M0, N0, lhs_tile, rhs_tile, {{dst}});
)_";
        }
    code +=
    // Calculate the destination indirect Y
R"_(
    LOOP_UNROLLING(int, i, 0, 1, M0,
    {
        g_dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)({{out}}_w * {{out}}_h) - 1);
        g_dst_indirect_y[i].v += g_ind_2 * (int)({{out}}_w * {{out}}_h);
    })
    }
    //------------------ END KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
)_";
    }

    else // non-root
    {
        code =
R"_(
    //------------------ START KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
)_"
    // IN_0/Out(Accumulator)   {{acc}}
    // IN_1(Operand)        {{operand}}
    // acc = operand + acc (mix-precision, broadcast, boundary aware)
R"_(
    {
        TILE(DATA_TYPE, M0, N0, operand_tile);
        T_LOAD({{DATA_TYPE}}, {{rhs_m0}}, {{rhs_n0}}, BUFFER, {{operand}}, {{rhs_start_ind_0}}, {{rhs_start_ind_1}}, 1, {{operand}}_stride_y, operand_tile);
)_";

        if(is_broadcast)
        {
            code +=
R"_(
        T_ELTWISE_BROADCAST_{{ELTWISE_OP}}_X({{DATA_TYPE}}, M0, N0, {{acc}}, operand_tile, {{acc}});
)_";
        }
        else
        {
            code +=
R"_(
        T_ELTWISE_{{ELTWISE_OP}}({{DATA_TYPE}}, M0, N0, {{acc}}, operand_tile, {{acc}});
)_";
        }
    code +=
R"_(
    }
    //------------------ END KERNEL {{meta_kernel_id}} ELTWISE_OP ---------------------
)_";
    }

    return code;
}

void ClTemplateElementwiseBinary::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
{
    vtable.declare_variable(
        _lhs,
        GpuKernelArgumentInfo(common_tensor_type),
        comp_group.is_intermediate_tensor(_lhs),
        "lhs");

    vtable.declare_variable(
        _rhs,
        GpuKernelArgumentInfo(common_tensor_type),
        comp_group.is_intermediate_tensor(_rhs),
        "rhs");

    vtable.declare_variable(
        _dst,
        GpuKernelArgumentInfo(common_tensor_type),
        comp_group.is_intermediate_tensor(_dst),
        "dst");
}

TagLUT ClTemplateElementwiseBinary::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
{
    TagLUT             lut{};
    const ITensorInfo *accumulator = _lhs;
    const ITensorInfo *operand     = _rhs;

    // Local build options
    lut["meta_kernel_id"] = id();
    lut["DATA_TYPE"]      = get_cl_type_from_data_type(_lhs->data_type());
    // Arguments and global shared variables
    const bool is_root = (comp_group.get_root_component()->id() == this->id());
    if(is_root)
    {
        lut["lhs"] = vtable.get_variable(_lhs);
        lut["rhs"] = vtable.get_variable(_rhs);
        lut["dst"] = vtable.get_variable(_dst);
        lut["out"] = vtable.get_variable(comp_group.get_any_dst_tensor());
    }
    else
    {
        // Determine which tensor is the accumulator
        if(comp_group.is_intermediate_tensor(_lhs))
        {
            accumulator = _lhs;
            operand     = _rhs;
        }
        else if(comp_group.is_intermediate_tensor(_rhs))
        {
            accumulator = _rhs;
            operand     = _lhs;
        }
        else
        {
            ARM_COMPUTE_ERROR("Invalid elementwise component linking");
        }
        lut["acc"]     = vtable.get_variable(accumulator);
        lut["operand"] = vtable.get_variable(operand);
    }
    switch(_attributes.operation())
    {
        case Attributes::ElementwiseOp::ADD:
            lut["ELTWISE_OP"] = "ADD";
            break;
        default:
            ARM_COMPUTE_ERROR("Arithmetic Operation not supported");
    }
    ARM_COMPUTE_ERROR_ON_MSG(detail::have_different_dimensions(accumulator->tensor_shape(), _dst->tensor_shape(), 0), "Only the operand can be broadcast to match the accumulator's shape");
    const bool is_broadcast = (operand->tensor_shape() != _dst->tensor_shape());

    // Set broadcast parameters
    // PRE: All tensors are broadcast-compatible
    if(is_broadcast)
    {
        // Note that n0 maps to input tensor dimension 0, m0 maps to input dimensions 1 and 2 because of our collapse strategy
        if(operand->dimension(0) == 1U && operand->dimension(1) == 1U && operand->dimension(2) == 1U) // Broadcast in X, Y, Z: collapsed rhs win [M0xN0] = [1x1]
        {
            lut["rhs_m0"]          = "1";
            lut["rhs_n0"]          = "1";
            lut["rhs_start_ind_1"] = "0";
            lut["rhs_start_ind_0"] = "0";
        }
        else if(operand->dimension(1) == 1U && operand->dimension(2) == 1U) // Broadcast in Y and Z: collapsed rhs win [M0xN0] = [1xN]
        {
            lut["rhs_m0"]          = "1";
            lut["rhs_n0"]          = "N0";
            lut["rhs_start_ind_1"] = "0";
            lut["rhs_start_ind_0"] = "g_ind_0";
        }
        else
        {
            ARM_COMPUTE_ERROR("Only support rhs broadcasting in all X, Y, Z dimensions, or just in Y and Z dimensions");
        }
    }
    else
    {
        lut["rhs_m0"]          = "M0";
        lut["rhs_n0"]          = "N0";
        lut["rhs_start_ind_1"] = "g_ind_1";
        lut["rhs_start_ind_0"] = "g_ind_0";
    }
    return lut;
}

CLBuildOptions ClTemplateElementwiseBinary::get_build_options(const ComponentGroup &comp_group) const
{
    CLBuildOptions build_opts{};
    /// NOTE: For now tile sizes (n0, m0) are set by the execution window. This may change in the future
    const auto         root_window      = comp_group.get_root_component()->template_writer()->get_window();
    const unsigned int n0               = root_window.x().step();
    const unsigned int m0               = root_window.y().step();
    const unsigned int partial_store_n0 = _dst->dimension(0) % n0;

    build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
    build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_lhs->data_type()));
    build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));

    return build_opts;
}

std::string ClTemplateElementwiseBinary::get_config_id() const
{
    std::string config_id{};
    config_id += lower_string(string_from_data_type(_dst->data_type()));
    config_id += "_";
    config_id += support::cpp11::to_string(_dst->dimension(0));
    config_id += "_";
    config_id += support::cpp11::to_string(_dst->dimension(1));
    config_id += "_";
    config_id += lower_string(string_from_data_layout(_dst->data_layout()));

    return config_id;
}

std::set<std::string> ClTemplateElementwiseBinary::get_headers_list() const
{
    return std::set<std::string>{ "helpers.h", "tile_helpers.h" };
}

Window ClTemplateElementwiseBinary::get_window() const
{
    ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");

    TensorShape output_shape = _dst->tensor_shape();
    // Collapse Dim 1 (W) and Dim 2 (H) together, leave Dim 0 (C) and upper dimensions unchanged
    // This is in line with the collapsing convention used by operators like Conv2d
    output_shape.collapse(2U, 1U);
    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / _dst->element_size(), _dst->dimension(0));
    Window             win                               = calculate_max_window(output_shape, Steps(num_elems_processed_per_iteration));

    return win;
}

} // namespace dynamic_fusion
} // namespace experimental
} // namespace arm_compute